Thomas Haynes and Sandip Seno
We propose a multiagent case-based learning (MCBL) framework in which agents learn cases to over-ride default behavioral rules. When the actual outcome of the action of an agent using its behavioral rules is not consistent with the expected outcome based on the model the agent has of other agents, the agent recognizes that a conflict has occurred and that its behavior is not appropriate in that situation. For those situations, the agent learns exceptions to its behavioral rules that are likely to prevent future conflicts. Agents follow their behavioral rules except when a learned case suggests alternative actions. Through this process, the agents dynamically evolve a behavior that is suited for the group in which it is placed.